The investment case for AI infrastructure now rests on a single shortage. Kindred Ventures has raised $355 million, and its founder Steve Jang points to a gap between compute demand and supply that could reach 60 gigawatts by 2030.
That gap is the thesis. Where demand outruns supply, early money flows.
Where the fund is looking
Kindred splits its focus three ways: AI applications, AI infrastructure, and physical AI covering autonomy, self-driving and humanoid robots.
One infrastructure holding is Architect Labs, a custom-chip lab letting companies co-design silicon with its team. The pitch is speed, giving robotics and autonomous-vehicle firms bespoke chips that cost, talent and tooling had kept out of reach. Front-end design alone often runs a year or two, before a multi-year path to production through fabs such as TSMC.
Token quality divide
The more useful idea Jang raises concerns tokens, the units of AI output. Not all of them carry equal value.
High-quality tokens come from reasoning models like Opus and the top GPT systems. They are expensive and suited to high-value work. Mid-quality tokens come from open-source models, cost less, and serve routine steps in an agent's workflow.
Why "token maxing" can mislead
This reframes a common assumption. Heavy token use is treated as a sign of demand, but volume alone says little.
The question is whether those tokens are high-quality and needed, or low-quality and wasted. Firms such as Cursor and Perplexity take cheaper open-source models, fine-tune them on proprietary data, and build feedback loops that lift quality without paying frontier prices.
Gap that is closing
The shift matters because the quality distance is shrinking. Open-source models once trailed the expensive ones by about six months.
Recent data shows that lag narrowing. As mid-tier models improve, the revenue case for premium-priced frontier models gets harder, and that reshapes the economics of every startup built on top of them.
Inference pyramid
Underneath sits a restructuring of who supplies compute. Three tiers have formed: hyperscalers serving frontier labs, neo clouds running their own data centres, and inference platforms such as Fal and Fireworks serving startups.
Neo clouds are being pushed out of hyperscaler facilities and carrying their own capital costs. Asset-light players like Hydra Host are stitching together spare capacity from greenfield sites, conversions and former crypto miners.
Jang's fund is a wager on scarcity. The token economy beneath it suggests the scarcity that pays off may not be raw compute, but the compute worth using.
